prashant retweeted
Hi, I’m Vibhor, Product Lead for Ecosystem at @OpenAI. I am going to use this space to share what we are doing to further the Ecosystem for our customers, and partners. DevDay feels like a good time to start.
One of the best parts of DevDay was the deep collaboration with partners: @Adobe, @Atlassian, @Canva, @Figma, @Instacart, @MagicPathAI, @Shopify, Superpower, and @tldraw.
While I went in excited I left with even more conviction in what we’re building together for our shared customers. It was also incredible to meet AI native partners and hear how our platform is empowering them build a sustainable business on ChatGPT 🎉
A conversation is often just the beginning of a task. You describe an idea, work through it, and eventually want something you can use: an edited image, a working prototype, a diagram you can keep refining. Getting there takes great tools and the people who know how to build them. Our ambition is to make those capabilities feel like a natural part of working with ChatGPT.
At DevDay, we shared several steps in that direction.
1. Plugin Extensions help our partners bring interactive experiences to ChatGPT. Adobe is a good example. You can ask for an image edit, then adjust the details with hands-on controls. With tldraw, you and the agent can work on a shared canvas. MagicPath brings interactive prototypes into the conversation, with room to edit them yourself.
These examples make the idea much more tangible:
- Adobe: video.tv.adobe.com/v/3503940
- tldraw: x.lingyaoai.com/tldraw/status/21053958…
- MagicPath: x.lingyaoai.com/skirano/status/2104994…
- Figma: lnkd.in/gaqHXZjT
- Shopify: lnkd.in/ghEXY5GT
2. A useful plugin can package skills, an MCP server, or both. That gives builders different ways to contribute: reusable workflow knowledge, access to live information, or actions that help someone finish a task. We introduced Plugin Creator and improvements to submissions, including automated checks and clearer feedback. Sites can also host supported plugins. There’s still work to do to make building and maintaining a plugin consistently straightforward.
3. We also announced improvements to plugin discovery, ranking and recommendations, in the directory and in conversations. This matters for both sides of the ecosystem. Users need to find something useful when they need it. Developers need a way for their work to reach those users. The goal is to make suggestions genuinely helpful, with good judgment about when they add value and when they get in the way.
4. MCP Events lets users ask ChatGPT to watch for supported events, such as a new message or a document comment, and respond when they arrive. It opens up another way for developers to make their tools useful beyond a single exchange.
5. ChatGPT Sites can now host MCP servers including plugin extensions! Create an MCP server with extensions, deploy the MCP server to Sites, turn the MCP server into a plugin, and install the plugin across platforms.
There’s still plenty to do, especially around reliability, discovery and the developer experience. Thank you to the teams and partners doing this work with us. The future feels bright, and we’re just getting started. Onward!
Plugin extensions let you build interactive panels, file viewers, and a sidebar home in ChatGPT.
Get clearer feedback when submitting your plugin for review.
With support for the proposed MCP Events specification, events in connected apps can start automations in ChatGPT:
developers.openai.com/plugin…
50
49
724
148,332
great opportunity to get your sign-in with chat questions answered!
I am going to try and answer a bunch of questions that people have raised about Sign in with ChatGPT in the next couple of days. Before that, I want to let you all in to why we shipped Sign in with ChatGPT (SIWC) in the first place. Clarify our intentions here before getting to the specifics.
191
prashant retweeted
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:
Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:
Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:
Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:
Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!
In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
1,434
5,933
51,165
6,502,489
prashant retweeted
This has transformed how we work at Clay. Self-serve analytics is finally real. Had a lot of fun building monty & excited to share how we did it!
One of our biggest bottlenecks was the queue between a good question and someone with the time and context to give a good answer. So our data team built Monty, a custom AI analyst who lives in Slack.
11
3
32
2,804
another devday in the books.
the highlight for me was meeting api and codex builders from cleveland, north carolina, michigan, and poland. many started using our products to build businesses or grow existing ones, with the simple goal of improving the lives of people they love.
now they’re shipping side projects, bringing other builders along, and fostering communities of hackers and innovators in their hometowns. trailblazers in their own right.
it reinforced something we talk about a lot on our team: we’ll never have the good fortune of meeting the vast majority of the people building with our products.
building for scale isn’t optional. it’s how we empower all the builders we’ll never meet.
10
477
perhaps expectedly, the coding harnesses from the labs have avoided semsearch, probably because they understand how quickly better models can make specialized retrieval optimizations obsolete.
282
gotta bring the right tool for the job!
i saw this coming while working on coding evals late last year. every team building coding agents was looking for a moat, and proprietary code search looked like a pretty compelling candidate.
it was an alluring story. a right-sized problem to tackle without a full-blown post-training team. lots of knobs to fiddle with, cheap ablations, and a plausible hill to climb.
but the bitter lesson struck again; those code search gains got washed away by RL scaling, just as my Q4 2025 was washed away asking customers to explain their code search so we could help them eval it. 😅
(turbopuffer co-founder) yeah, Cursor called us in July and told us their new model code evals didn't show the same gain from semsearch as previous models. selecting the right piece of context in a single turn used to be huge, but modern code-local agents can answer complex questions with simple search tools like grep(1)
in local code bases, with a low-scale amount of data to search, and when you need a full copy of the data on your machine anyway (to edit and run), offloading to search infrastructure is overkill. grep works fine, and is how it's been done by expert programmers for decades
search engines are necessary if you have lots of data to search over, and want to select the right context for the task rather than downloading all of it to your agent's disk. search engines also let you spend compute up front to save turns at search time. large datasets or high qps are where it shines
we are close to the Cursor team, and will work with them on more large-scale search in the future :) we'd have done the same thing
2
5
927
this tuesday will be my third DevDay.
there’s something special about how the whole company rows in the same direction in the weeks leading up to it. then you spend the day with the people who built the models and products, and the developers building on top of them, and it brings you back to why any of this matters. bringing safe AGI to everyone is a lot more tangible when you see what people are doing with it.
the show returns Tuesday. biggest one yet, in more ways than one :)
join the livestream devday.openai.com or you catch the recording later on YouTube
3
1
22
1,312
i’m building a pretty ridiculous team to work with startups at OpenAI.
you’ll get to work with ambitious founders on hard technical problems alongside teammates who make you better. and you’ll be customer zero for all the new stuff our research and engineering teams are cooking.
direct experience building and optimizing agents is required. bonus points if you’ve worked in sales or solutions architecture before.
apply here — openai.com/careers/applied-a…
39
25
678
62,254
codex has a familiar shape: open source app server you can fork, a desktop app that adds a polished local experience, and Agents API for the “please run and scale this for me” crowd. we’re already seeing startups fork the app server and build pretty interesting products on top.
very reminiscent of open core cos like mongodb / confluent.
claude has local and managed offerings too, but no equivalent open source core where you can fork the harness and get weird with it.
this used to sustain an entire tech discourse cycle. now by the time you’ve typed “open source vs closed source,” three new models have shipped and your agent has picked that in-distribution dependency for you. 🤷♂️
3
1
17
1,894
presales engineering for agents is getting interesting because its getting harder to prove the work is good.
the agent cooked for 4.5 hours and produced something that takes real expertise to judge. even if the prospect has that expertise, reviewing it is work.
some cyber products have a compelling demo: find vulnerabilities in the prospect’s surface area and show them the findings. a concrete claim they can check, about something they care about.
what’s the equivalent in fintech, healthcare, or other verticals? something a prospect can validate during the pitch without having to redo the agent’s work.
3
1
11
1,051
prashant retweeted
GPT-6 Luna is half the price of 5.6 Luna, which was already an astonishingly cheap model given how capable it is
Luna is my favorite model for building product features thanks to its cost (and speed)
Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe.
GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale.
We’ve also made caching and inference more efficient, and we’re passing the savings directly to you: 50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing.
143
58
1,642
121,508